BotPsy tests compact agents for private local assistants
Tether AI Research released an open 0.8B tool-using model designed for on-device assistants, reporting strong benchmark gains but limited scope.
November 10, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- AIGC Tripo showed a stack combining Tripo, Fable 5.2, Blender, and Claude to build an open-world RPG. X / Tripo AI ↗
- Models Mistral used Voxtral Transcribe 3 to transcribe every main-stage talk at AI Engineer Paris in real time. X / Mistral AI ↗
- Research Google DeepMind said Gemini agents generated more than six million Lean proof lines, leaving humans to review the critical one percent. X / Google DeepMind ↗
- Models Tencent Hunyuan announced that its Hy4 model is available through OpenRouter. X / Tencent Hunyuan ↗
- Agents Cloudflare introduced a Web Search API designed for AI applications. InfoQ 中文 ↗
- AIGC Hailuo AI promoted H3 Image for gear-focused image generation and editing. X / Hailuo AI ↗
- Funding A Latent Space report said Mercor raised $750 million at a $20 billion valuation and reached a $3 billion run rate. Latent Space ↗
- Industry Tencent Hunyuan also made a Hy4 API live on Atlas Cloud. X / Tencent Hunyuan ↗
- Industry Sakana AI appointed retired Admiral Michael Rogers and former CIA officer Tom Moran as senior advisers. X / Sakana AI Labs ↗
- Open Source MBZUAI’s Institute for Foundational Models released PAN 2, an open 29-billion-parameter world model for physical AI. MarkTechPost ↗
- Agents MiniMax demonstrated an H3 agent directing cinematic video through HeyGen’s MCP integration. X / MiniMax AI ↗
Human-agent teams run auditable X-ray experiments
A new arXiv study documents autonomous synchrotron experiments, measuring how humans and coding agents share decisions, errors and scientific work.
- ⚡ Quick hits
- AIGC Fish Audio released Live 3, a full-duplex voice model with a reported 186-millisecond reaction time. MarkTechPost ↗
- Industry A Qbit report showed an autonomous tractor performing deep tillage exceeding 30 centimeters in Xinjiang fields. 量子位 ↗
- AIGC MiniMax said H3 Image supports 16 references and produces outputs two to three times faster. X / MiniMax AI ↗
- Models Alibaba’s Qwen team teased Qwen4 with stronger knowledge, world understanding, coding, and agentic capabilities. X / Alibaba Qwen ↗
- Models AWS added Kimi K3.1 Pro to Amazon Bedrock for enterprise model access. AWS Machine Learning ↗
- Agents ElevenLabs announced Routines for ElevenAgents, enabling multi-step work before, during, and after customer conversations. ElevenLabs ↗
- Agents AWS published a LangGraph and Bedrock AgentCore tutorial for building a daily dependency watcher with web search. AWS Machine Learning ↗
- Research Meta FAIR released e-means, a FAISS-integrated clustering algorithm that reaches Lloyd k-means quality in 1.4–2.6 epochs. MarkTechPost ↗
Microsoft finds dormant-neuron backdoor in AI agents
Microsoft researchers show that fewer than ten dormant neurons can trigger malicious agent behavior without retraining the underlying model.
- ⚡ Quick hits
- AIGC Canva made its Craft generative design capability available through an API for developers to integrate into their own products. Canva ↗
- Funding Mercor raised $750 million at a $20 billion valuation to expand custom AI-agent training and evaluation from frontier labs to enterprises. SiliconANGLE ↗
- Open Source Genspark and Baseten said their post-trained open-weight model, based on Qwen3.5-397B, beats Opus 5.5 on deep-research tasks. Baseten ↗
- Research Goodfire released natural-language descriptions for 1,500 neurons in Qwen3.6 35B-A3B after agent-driven interpretability investigations. Goodfire AI ↗
- AIGC ElevenLabs released Scribe v3 and Scribe v3 Live, adding real-time transcription and diarization for ElevenAgents and its API. ElevenLabs ↗
- Agents OpenAI launched Business Hub beta, letting verified companies manage ChatGPT profiles and publish business agents for customer discovery. OpenAI ↗
- AIGC Luma launched an MCP integration for ChatGPT and Claude, allowing users to build consistent character campaigns inside conversations. Luma AI ↗
Anthropic expands Claude into collaborative office work
Claude gains Whiteboards, Sheets and Data tools as Anthropic pushes beyond coding into shared planning, analysis and workplace execution.
- ⚡ Quick hits
- AIGC Adobe previewed Firefly Design, a cross-medium workspace spanning vector, raster, motion, 3D, and print workflows. Adobe Firefly ↗
- Industry Pinecone added OpenAI text-embedding-3 models to its integrated embedding feature, allowing plain-text fields to generate vectors directly. Pinecone ↗
- AIGC Runway showcased real-time video prototypes that animate sketches, rearrange scenes, and generate context-aware video inside an interface. Runway ↗
- Safety NVIDIA released two Nemotron synthetic datasets for Japanese personas and personal-information detection through Hugging Face. NVIDIA AI ↗
- AIGC Meshy launched an arcade of 18 AI-built games and showed a playable kart game created without human-written code. Meshy ↗
Google releases Gemini 4 Carbon for long-horizon agents
Google’s new frontier model targets extended coding, scientific, cybersecurity and knowledge-work tasks with a 1-million-token context.
- ⚡ Quick hits
- Open Source Zyphra released ZONOS3-Mini, an Apache-licensed 800M-parameter TTS model with about 100M active parameters for local real-time use. MarkTechPost ↗
DeepMind upgrades Gemini Robotics and opens safety benchmark
Gemini Robotics 2.5 expands robot dexterity while DeepMind releases ASIMOV-Jailbreak to test unsafe instructions across varied physical scenes.